The wording of a prompt used to be most of the work. As models got better at understanding intent, the difficulty moved somewhere else: choosing what the model sees.
By OneShopAI · · 4 min read
Two years of collected prompt tricks have quietly stopped mattering. Offering a model a tip, insisting it is a world-class expert, threatening it with consequences — these were real effects on weaker models, and on current ones they mostly do nothing measurable. The models got better at inferring what you meant, which retired an entire genre of advice.
What did not retire is specificity. "Summarise this" still produces a guess about length, audience and emphasis, because those are genuinely unspecified and no model can read your mind. The difference is that being specific was never a trick. It is just describing the job.
Look at where time actually goes on a working AI feature now and almost none of it is wording. It goes on retrieval: which documents come back for this question, how they were chunked, whether the right one ranks in the top three. It goes on history: how much of this conversation to keep, what to summarise, what to drop. It goes on tool results: how much of that API response the model needs versus how much is padding the bill.
All of those are the same decision in different clothes — what earns space in the context window. That is a resource-allocation problem, not a writing one, and it is why the work is increasingly called context engineering.
The intuitive response to a million-token window is to stop choosing. Put everything in; let the model sort it out. In practice this goes badly on three fronts at once. Cost scales with what you send, on every call. Latency scales with it too, because time to first token is dominated by how much input has to be read. And attention is not uniform across a very long context — material buried in the middle is measurably more likely to be missed than the same material near the start or end.
A large window changes the economics of retrieval. It does not remove the need for it. The teams getting the most out of long-context models are using the space deliberately, not filling it.
If you are picking up this skill now, the leverage is not in collecting prompt patterns. It is in understanding what the context window is, what retrieval does, and how to tell whether the right information reached the model at all — which is a debugging skill more than a writing one.
A useful habit: when output disappoints, ask first whether the model had what it needed. Most of the time the answer is no, and the fix is upstream of the prompt entirely.